The Agent Skills Directory

[COMMAND_EXECUTION]: Potential shell command injection in Step 1. The skill instructs the agent to interpolate a user-provided YouTube URL directly into a python3 command string. If the URL contains shell metacharacters like backticks or semicolons, it could allow for arbitrary command execution on the host system.
[COMMAND_EXECUTION]: Unsafe command construction in Step 5. The skill uses AI-generated content (summary title and guest name) to build arguments for the save_to_wiki.py script. Maliciously crafted podcast content could result in a summary that triggers shell injection when the command is executed.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests untrusted data from YouTube transcripts and uses the processed results to drive downstream actions.
Ingestion points: YouTube transcripts processed in Step 1 (via get_transcript.py).
Boundary markers: Absent; there are no instructions to differentiate between podcast content and potential embedded commands.
Capability inventory: File system writing to local Obsidian directories, shell command execution for Feishu updates, and data submission to external services.
Sanitization: No sanitization or escaping of the transcript content is mentioned before it is processed or used in shell commands.
[EXTERNAL_DOWNLOADS]: The skill requires the agent to interact with and upload content to unverified external services, specifically https://madopic.thus.chat for image conversion and https://zara.faces.site/ai as a data source.
[CREDENTIALS_UNSAFE]: Exposure of internal system identifiers. The Feishu parent node token TOSJwKzxTiFdiRk0aducHNBFntg is hardcoded in the instructions, exposing the structure of the user's specific knowledge base environment.

read-podcast-workflow